A Vision-Based Approach for Multi-Component Pasture Biomass Estimation

Abstract

Accurate estimation of grassland biomass is fundamental for designing sustainable grazing strategies and optimizing pasture management, yet conventional field methods remain labor-intensive, destructive, and difficult to scale. In this study, we exploit recent advances in computer vision to estimate multiple components of grassland biomass from overhead RGB imagery. The analysis is based on the Image2Biomass dataset, comprising 1162 annotated images of grasslands across Australia, each paired with laboratory-validated biomass measurements. A structured preprocessing pipeline was implemented, including exploratory data analysis, outlier mitigation, and logarithmic transformation of target variables, in accordance with the dataset evaluation protocol. Building on this foundation, we propose an encoder–decoder regression framework that integrates self-supervised visual representation learning with ensemble-based prediction. The encoder employs a DINOv2 Giant model as a feature extractor to capture detailed spatial and structural characteristics of the sward, while the decoder uses a stacking ensemble combining LightGBM, XGBoost, and Ridge Regression. Across 15 repetitions of shuffled four-fold cross-validation, the cross-fitted stacking ensemble achieved a weighted coefficient of determination of Rw2=0.7757±0.0171, a weighted mean absolute error of 8.3867±0.2736 g, and a weighted root mean squared error of 13.3994±0.4995 g. The ensemble significantly outperformed LightGBM, XGBoost, and Ridge Regression on the primary weighted R2 metric in paired comparisons (Holm-adjusted p<0.001 for all three comparisons). These results highlight the potential of computer vision methods as scalable, non-destructive tools for operational monitoring of grassland biomass, supporting more informed agronomic decision-making in pasture-based systems.